Architecture · Rack · Simulation
Rack power density & capacity planning
A site has a fixed amount of power. How many racks it can run depends on how much each rack draws, how efficient the building is, what is held in reserve and how much heat the cooling can remove. Change any of them and watch the plan move.
Step 1 · Rack architecture
Choose a rack
Each figure says what kind of number it is. A maximum rating, an analyst’s estimate, a rated TDP and a peak are not the same, and the planner never mixes them without saying so.
NVIDIA GB300 NVL72
Vendor figure- Architecture
- 72 Blackwell Ultra GPUs and 36 Grace CPUs on 18 compute trays, one NVLink domain
- GPUs
- 72
- IT power per rack
- 135 kW · rated (TDP); up to 155 kW at peak
- Heat outputIT heat load: Almost all the electricity IT equipment uses becomes heat, so the IT heat load in kilowatts thermal (kWt) is about equal to the IT power in kW.
- ≈ 135 kWt per rack
- Cooling
- Hybrid: liquid to the compute and switch trays, air for the power shelves; a CDU in the row
- Electrical
- Eight power shelves with energy storage that smooths the load swings of training; the electrical design is sized to the peak, not the rating
Lenovo’s build guide: “Each rack consumes 135 kW TDP; up to 155 kW peak depending on workload.” Other builders quote 132–140 kW. The 5 MW exercise uses 135 kW as an illustrative figure.
Source: Lenovo Press, Lenovo NVIDIA GB300 NVL72 Rack Scale AI — Product Guide (LP2357)
Planning on 135 kW per rack.
Step 2 · The site’s budget
Capacity plan
Power redundancyRedundancy: How much electrical and cooling equipment is installed beyond what the load needs. It changes what is installed, not what the racks draw.
Total IT power
2.7 MW
20 × 135 kW
Facility power
3.38 MW
IT × PUE 1.25
Usable IT capacity
3.6 MW
4.5 MW ÷ 1.25
Maximum racks
26
Power-limited (cooling allows 29)
IT heat loadTons of refrigeration: Tons of refrigeration, the unit chillers are often rated in: 1 ton = 3.517 kWt.
2.7 MWt
≈ 768 tons of refrigeration
Power used
75%
of usable IT capacity
Power headroom
900 kW
room for 6 more racks
Cooling headroom
1.3 MWt
68% of cooling used
Facility power: 5 MW available
- IT load 2.7 MW
- Facility overhead (PUE) 675 kW
- Unused 1.13 MW
- Reserve 500 kW
Cooling: 4 MWt available
- IT heat 2.7 MW
- Spare cooling 1.3 MW
- Within budget (20)
- Room for more (6)
What N+1 means for installed equipment
UPS modules (500 kW each, illustrative)
6 needed, 7 installed: 3.5 MW of installed capacity for an operating load of 2.7 MW.
Cooling units (500 kWt each, illustrative)
6 needed, 7 installed: 3.5 MWt of installed capacity for an operating load of 2.7 MWt.
Redundancy changes what is installed, not what the racks consume. In 2N each side normally carries about half the load and must be able to carry all of it, so the installed capacity doubles while the operating draw stays at the IT load times the PUE (plus small losses from running equipment part-loaded). That is why the planner reads facility power as firm capacity: switching from N+1 to 2N changes the equipment list and the cost, not the number of racks.
Step 3 · Challenge
The 5 MW AI data centre
You are responsible for capacity planning at a 5 MW AI data centre. The facility has:
- 5 MW total facility power budget
- PUE of 1.25
- 10% reserved facility capacity
- NVIDIA GB300 NVL72 racks
- Illustrative rack IT power of 135 kW
How many GPU racks can the facility support within its usable power budget?
A power-budget result, not a build plan. 26 racks fit the power budget; that does not prove 26 can be installed. Feasibility also depends on cooling distribution, electrical topology, rack space, floor loading, network capacity and the equipment’s own specifications.
Step 4 · Density
Same budget, different racks
Racks the current budget supports (5 MW, PUE 1.25, 10% reserve), power only.
- CPU rack8 kW450CPU only
- DGX H100 ×440.8 kW882,816 GPUs
- GB200 NVL72120 kW302,160 GPUs
- GB300 NVL72135 kW261,872 GPUs
- Vera Rubin NVL72180 kW, assumed201,440 GPUs
Higher rack density means more compute per rack, but also more power and cooling per rack. The same budget holds far fewer racks of a denser kind. That is not automatically better: it moves the problem to liquid cooling, heavier floors, new power distribution and a bigger blast radius when one rack fails. Whether it lowers total cost depends on the site, the workload and the price of the equipment, not on density alone. The GPU counts are not comparable as compute either: an H100 and a Blackwell Ultra GPU do very different amounts of work.